support fma

[yuriisig]

when optimization by means of fma of instructions is realized?

[TimP]

a) target architecture includes fma, e.g. -mavx2, -xHost on avx2 platform, -mmic, -arch:AVX2 (or past IA64, PPC, mips)

b) floating point multiply-add sequence e.g. a*b+c, c-a*b

c) compiler may analyze latencies; e.g. IA dot product needs higher degree of riffling to maximize fma performance

[yuriisig]

Points a) and b) I know. I didn't understand point c). The compiler has command /Qfma, but it doesn't influence the generated code. My processor Intel® Core™ i7-5820K

.

[JVanB]

I always use /QxCORE-AVX2 to get ifort to output FMA instructions for my Core i7-4770.

[yuriisig]

>I always use /QxCORE-AVX2 to get ifort to output FMA instructions for my Core i7-4770.

You watched an assembly code? There are fma commands?

[JVanB]

Most definitely /QxCORE-AXV2 generated assembly code with FMA instructions. I have posted disassemblies lately in this forum.

[yuriisig]

>Most definitely /QxCORE-AXV2 generated assembly code with FMA instructions.

I passed the test: the FMA commands appeared. Thanks.

[TimP]

/Qfma will have no effect in ifort unless core-avx2 target is set (often done by QxHost).  Then it will be set by default; /Qfma- might be used to prevent use of fma.  As I mentioned, ifort might still choose not to use fma in situations where the longer latency will reduce performance.

I don't know why ifort would not accept /arch:AVX2 when some other compilers require that rather than /arch:CORE-AVX2.